def check_hat(args):
""" Helper function that handles hat manipulations, like flip and scale """
try:
folder = get_imgs("crimmis_hats/")
hat = Image.open(folder.get('0'))
for arg in args:
if arg.startswith('type='):
value = arg.split('=')[1]
hat = Image.open(folder.get(value))
w_offset, h_offset = 150, 0
hat_width, hat_height = 350, 300
for arg in args:
if arg == 'flip':
hat = hat.transpose(Image.FLIP_LEFT_RIGHT)
w_offset, h_offset = 0, 0
if arg.startswith('scale='):
value = float(arg.split('=')[1])
hat_width, hat_height = int(hat.width * value), int(hat.height * value)
hat = hat.resize((hat_width, hat_height))
except:
return None, None, None
return hat, w_offset, h_offset
def train(n_clusters=N_CLUSTER):
imgs = get_imgs("database/")
sift = cv2.xfeatures2d.SIFT_create()
desc_list = []
for word, img_list in imgs.items():
for img in img_list:
kp, desc = sift.detectAndCompute(img, None)
desc_list.extend(desc)
kmeans = MiniBatchKMeans(n_clusters=n_clusters, n_init=1)
kmeans.fit(desc_list)
dump(kmeans, f'kmeans{n_clusters}.joblib')
dataset=FLAGS.dataset
img_size= (640,640) if dataset=='DRIVE' else (720,720) # (h,w) [original img size => DRIVE : (584, 565), STARE : (605,700) ]
img_out_dir="{}/segmentation_results_{}_{}".format(FLAGS.dataset,FLAGS.discriminator,FLAGS.ratio_gan2seg)
model_out_dir="{}/model_{}_{}".format(FLAGS.dataset,FLAGS.discriminator,FLAGS.ratio_gan2seg)
auc_out_dir="{}/auc_{}_{}".format(FLAGS.dataset,FLAGS.discriminator,FLAGS.ratio_gan2seg)
train_dir="../data/{}/training/".format(dataset)
test_dir="../data/{}/test/".format(dataset)
if not os.path.isdir(img_out_dir):
os.makedirs(img_out_dir)
if not os.path.isdir(model_out_dir):
os.makedirs(model_out_dir)
if not os.path.isdir(auc_out_dir):
os.makedirs(auc_out_dir)
# set training and validation dataset
train_imgs, train_vessels =utils.get_imgs(train_dir, augmentation=True, img_size=img_size, dataset=dataset)
train_vessels=np.expand_dims(train_vessels, axis=3)
n_all_imgs=train_imgs.shape[0]
n_train_imgs=int((1-val_ratio)*n_all_imgs)
train_indices=np.random.choice(n_all_imgs,n_train_imgs,replace=False)
train_batch_fetcher=utils.TrainBatchFetcher(train_imgs[train_indices,...], train_vessels[train_indices,...], batch_size)
val_imgs, val_vessels=train_imgs[np.delete(range(n_all_imgs),train_indices),...], train_vessels[np.delete(range(n_all_imgs),train_indices),...]
# set test dataset
test_imgs, test_vessels, test_masks=utils.get_imgs(test_dir, augmentation=False, img_size=img_size, dataset=dataset, mask=True)
# create networks
g = generator(img_size, n_filters_g)
if FLAGS.discriminator=='pixel':
d, d_out_shape = discriminator_pixel(img_size, n_filters_d,init_lr)
elif FLAGS.discriminator=='patch1':
d, d_out_shape = discriminator_patch1(img_size, n_filters_d,init_lr)
_, Y_train, _, _ = utils.load_data()
X_train = np.load(hog_train_imgs_file)
X_test = np.load(hog_test_imgs_file)
elif mode == 'train' or not os.path.exists(utils.train_imgs_file) or not os.path.exists(utils.train_labels_file) \
or not os.path.exists(utils.test_imgs_file) or not os.path.exists(utils.test_labels_file)\
or not os.path.exists(hog_train_imgs_file) or not os.path.exists(hog_test_imgs_file):
print('Train BEGIN')
utils.repartition_data(father_path=os.path.abspath(
os.path.join(os.getcwd(), "..")),
test_rate=0)
imgs_train, Y_train, _, _ = utils.load_data()
X_train = utils.get_hog_features(imgs_train)
# X_val = utils.get_hog_features(imgs_val)
imgs_test = utils.get_imgs([test_imgs_path],
max_pool=True,
homomorphic=False,
file_type_list=['.bmp', '.png'],
equalize=False,
morphology=False)
X_test = utils.get_hog_features(imgs_test) # 获得HOG特征向量
np.save(hog_train_imgs_file, X_train)
np.save(hog_test_imgs_file, X_test)
else:
raise Exception('Mode not defined')
if not os.path.exists('Q1/linear_svc.model') or mode == 'repartition':
linear_svc = svm.LinearSVC(loss='hinge', max_iter=1000)
linear_svc.fit(X_train, Y_train)
pickle.dump(linear_svc, open('Q1/linear_svc.model', 'wb'))
linear_svc = pickle.load(open('Q1/linear_svc.model', 'rb'))
pred_test = linear_svc.predict(X_test)
files_name = os.listdir(test_imgs_path)
基本策略：先用max_pooling缩小图片，用UNET预测。然后将预测得到的图片做线性插值，恢复为原来的样子。
"""
import sys
import os
import numpy as np
import UNet
import utils
import cv2
mode = sys.argv[1]
PREDICT = 'predict'
TRAIN = 'train'
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5)) # 形态学处理（膨胀、腐蚀等）所用的核
# 获取原始训练图像和掩码图像
imgs = utils.get_imgs(['init_imgs'], file_type_list=['.bmp', '.png'], homomorphic=True,
max_pool=False, morphology=True, bit_wise=True)
mask = utils.get_imgs(['mask_imgs'], file_type_list=['.bmp', '.png'], max_pool=False,
morphology=False, bit_wise=True)
# 预处理用于训练的输入图像和掩码图像并保存
if not os.path.exists('Q2'):
mode = TRAIN
os.mkdir('Q2')
if not os.path.exists('Q2/imgs'):
os.mkdir('Q2/imgs')
if not os.path.exists('Q2/mask_imgs'):
os.mkdir('Q2/mask_imgs')
for i in range(imgs.shape[0]):
tmp = imgs[i].astype(np.uint8)
tmp = utils.adjust_gamma(tmp, gamma=6)
tmp = cv2.erode(tmp, kernel)
from joblib import dump, load
from utils import get_imgs
import cv2
kmeans = load("kmeans100000.joblib")
imgs = get_imgs("database/")
sift = cv2.xfeatures2d.SIFT_create()
desc_list = []
for word, img_list in imgs.items():
for img in img_list:
kp, desc = sift.detectAndCompute(img, None)
desc_list.append(desc)
def eval_model(self):
device = self.device
model = self.model.to(device)
# model.load_state_dict(torch.load(self.path))
# model.eval()
dfcn_feats, dcnn_feats = [], []
names = list()
h5f1 = h5py.File('featsDCNN.h5', 'w')
h5f2 = h5py.File('featsDFCN.h5', 'w')
imgs = get_imgs('net/train')
train_data = OneData(config=self.config,
data_dir=self.train_dir,
transform=train_transform(self.config))
train_loader = DataLoader(dataset=train_data,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.config.num_workers)
# print(imgs)
for idx, (inputs, labels) in enumerate(train_loader):
# name = os.path.split(inputs)[-1]
# inputs = inputs[np.newaxis, :]
inputs = inputs.to(device)
# labels = labels.to(device)
# print(labels)
# print(inputs[0])
output, dfcn, dcnn = model(inputs)
# dfcn = np.array(dfcn)
# dfcn = np.squeeze(dfcn)
# dcnn = np.squeeze(dcnn)
# print(dfcn.shape)
# print(dcnn.shape)
for it in dfcn:
tmp = list()
for i in it:
tmp.append(float(i))
#
# it = list(it)
# it = np.squeeze(it)
# print(it.shape)
dfcn_feats.append(tmp)
for it in dcnn:
tmp = list()
for i in it:
tmp.append(float(i))
dcnn_feats.append(tmp)
# return
# dfcn_feats.append(dfcn)
# dcnn_feats.append(dcnn)
print("extracting feature from image No. %d , %d images in total" %
((idx + 1), len(train_loader)))
for name in labels:
names.append(name)
# names.append(name)
# if idx == 9:
# break
# print(dcnn_feats)
# print(np.array(dcnn_feats).shape)
feats = np.array(dcnn_feats)
print(feats)
print(feats.shape)
# return
h5f1.create_dataset('dataset_1', data=feats)
feats = np.array(dfcn_feats)
h5f2.create_dataset('dataset_1', data=feats)
names = [name.encode() for name in names]
h5f1.create_dataset('dataset_2', data=np.string_(names))
h5f2.create_dataset('dataset_2', data=np.string_(names))
h5f1.close()
h5f2.close()
